Information retrieval + RL

1. Ranking as Sequential Decision Making

　　Advantages: beyond independent relevance

 

2. RL: Learn to make good sequences of decisions

3. Alpha Go:

Breadth reduction: Policy Network 在某一手，下某些区域，是臭棋，可通过PN判断出来不必搜索，因而减少树的宽度。

Depth reduction: Value Network 在树搜索中，某一节点下注定赢不了棋，可通过VN剪枝，因而减少树的深度。

 

4. Ranking evaluation: NDCG (Normalized Discounted Cumulative Gain); Map(Mean average precision)

5. Monto-Carlo search